JD's catalog runs on no Nvidia

One of the world's largest retailers just described, in a 54-author paper, the machine that reads its store. JD's Oxygen AI Item Center covers tens of thousands of product categories and absorbs hundreds of millions of item updates a day — the language and vision models that decide what each of tens of billions of listings actually *is*. The buried detail: the whole pipeline runs on Huawei Ascend NPUs, China's domestic answer to the Nvidia chips it can't freely buy.

That last fact is the reason to read it. Ascend has mostly surfaced in lab-scale benchmark demos; here it is the operational backbone of a hyperscale retailer, running a frontier-adjacent production AI system with zero Nvidia in the loop. It is the most concrete public evidence yet that China's compute-sovereignty push is load-bearing today, not aspirational.

The engineering is deliberately un-flashy where it counts. JD calls the models 'self-evolving,' but the mechanism is modest and honest: incremental fine-tuning patches specific knowledge gaps while fighting the forgetting that usually follows, and the million-entry ontology is pushed out into a separate knowledge base so it can update continuously without retraining the model. JD reports 94.2% precision and auto-filling more than 80% of core attributes at listing time — its own production numbers, with no outside audit, as is normal for a systems paper. The claim that matters needs no audit: the chips underneath are Chinese.